[dmccoystephenson]

Summary

• baseline-write and fixture in gates/save_compat.py now require at least one file. Previously both tested bool(digests(...)), and digests() records directories as "dir", so a baseline write that left only an empty slot_1/, or a fixture archive holding only directories, passed with 0 file(s) in the detail. This is the second, smaller case of Gates can pass having checked nothing: every consumer skipped, or saves with no files #21.
• A file_count() helper replaces the two inline len([v for v ... if v != "dir"]) counts, so the number reported in the detail and the number the check tests are the same value.
• The module docstring and the README save-compatibility section state the new rule.
• New tests/test_save_compat_files.py covers file_count() over digests() output (empty slot directory, nothing written, mixed files and directories, an empty file).

Part of #21. The consumers half of #21 (every consumer skipped) is untouched: it needs the owner's choice between a new input and an unconditional control failure, so the issue stays open.

Verification

• python3 -m unittest tests.test_save_compat_files tests.test_save_compat_migrations tests.test_save_compat_view: 24 tests, OK (Python 3.8, the interpreter this dispatch can run). CI runs the full suite on 3.12.
• Mutation check: with file_count() changed to len(d) (the old entry-counting behaviour), 2 of the 4 new tests failed; with the file restored, all 4 passed.
• Local gate run: UNVERIFIED. python3.12 and env-prefixed gate commands require an approval this headless dispatch cannot receive, so no real save_compat.py run against tak was made and no PASS/FAIL assertion lines can be quoted. The unit tests exercise file_count() but not the two result.check(...) call sites themselves.

Test plan

• py_compile of the changed files; unit tests pass locally
• CI (Check / compile) green on the PR head
• Owner: a real save-compat run, e.g. REPOSITORY=Stephenson-Software/tak SHA=<tak main sha> BASELINE_REF=v0.2.0 SCENARIO=scenarios/tak_slots.py DISTRIBUTION=tak WORK=/tmp/prg-work python3.12 gates/save_compat.py, with baseline-write reading PASS ... N file(s) written for N > 0

Merge hold

This changes what an existing assertion (baseline-write, fixture) accepts, which is on the dev loop's do-not-auto-merge list, and the gate-run anchor could not run. The PR is therefore left open for owner review rather than merged. It is stricter only: no run that previously failed can now pass.

Skipped issues

• installed_commit reads a stderr line as the commit, failing candidate-install on a correct install #19 (installed_commit reading stderr): the fix is in gates/common.py, which is do-not-auto-merge, and also needs a gate run this dispatch cannot make.
• workflow_call: 'Checkout gates' appears to check out the caller, not this repository #24 (workflow_call checkout): needs an owner decision on whether workflow_call is supported, and a call from a second repository to verify.

This PR description was drafted during a Gardener session (https://github.com/Stephenson-Software/gardener).

🤖 Generated with Claude Code

---

drafted by Claude on behalf of Daniel Stephenson

[dmccoystephenson]

Self-review rubric (head c332c81):

• Scope: PASS — three files, all for Gates can pass having checked nothing: every consumer skipped, or saves with no files #21's save-data half: gates/save_compat.py (helper + two checks + docstring), README.md (one sentence), tests/test_save_compat_files.py (new).
• Tests-new: PASS — file_count() is exercised by four tests in tests/test_save_compat_files.py.
• Tests-fix: PASS, with a limit — the old behaviour was reproduced as a mutation (file_count → len(d), i.e. counting directories): 2 of 4 new tests failed; restored, 4/4 passed. The two result.check(...) call sites themselves are not unit-tested; only a real gate run reaches them.
• Sibling structure: PASS — the new test module follows test_save_compat_migrations.py / test_save_compat_view.py (docstring, sys.path insert, import save_compat, unittest.TestCase).
• Sibling renames: N/A — nothing renamed.
• Docs: PASS — Phase 7 rows checked: README assertion list (names and order unchanged), README result.json keys (files / fixture.files keep their meaning, a count of files), module docstring (baseline-write and fixture lines updated), save-compat.yml header (does not describe either assertion; unchanged), CLAUDE.md (no env var or command changed). No row needed a correction beyond the commit itself.
• Visuals: N/A — no README image.
• Issue resolution: PASS — the PR says "Part of Gates can pass having checked nothing: every consumer skipped, or saves with no files #21", not Closes; the consumers half remains open for an owner decision.
• CI + local gate run: UNVERIFIED — CI compile (py_compile + unittest discover on 3.12) is green on the head. The local save-compat run could not be made: python3.12 requires an approval this headless dispatch cannot receive.
• Gate-run evidence: FAIL — no PASS/FAIL assertion lines can be quoted, for the reason above.
• Input parity: N/A — no workflow touched.
• Read-only: PASS — git grep -n "secrets\." .github/workflows is empty; no workflow changed.
• Stdlib only: PASS — no new import in gates/.
• No tag moved: PASS — no tag operation was performed.

Merge hold: baseline-write and fixture change what they accept (stricter only), which is on the do-not-auto-merge list, and the gate-run anchor is UNVERIFIED. Left open for the owner, with a real tak save-compat run requested in the test plan.

This comment was drafted during a Gardener session (https://github.com/Stephenson-Software/gardener).

---

drafted by Claude on behalf of Daniel Stephenson